Unverified50% confidenceFactExact time
The paper gives a lower bound: after pretraining, the hallucination rate is at least equal to the proportion of facts in the training set that appear only once.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe closer an event is to the cutoff date, the smaller its share in the training corpus, and the less clearly the model has mastered it.75% similarVerifiedThe probability of hallucination increases significantly when contradictory information exists in the training data or when the model is asked about domains that are sparsely represented in the training data.74% similarVerifiedLarge models output the most likely next word rather than precise retrieval of facts, which leads to hallucinations in domains with sparse coverage in the training data.74% similarUnverifiedThe second term (1-M/N)(1-1/K) of the error rate lower bound corresponds to the traditional coverage problem, where the model can only guess randomly for facts that never appeared during training, with a correct-guess probability of 1/K.74% similarUnverifiedPretraining's footprint on VRAM (or NPU on-chip memory) is usually several times that of inference, and it is extremely sensitive to numerical precision; a single overflow or underflow during mixed-precision training can cause the loss to diverge.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063190API
curl https://kongchang.com/api/v1/knowledge/claims/1063190MCP
get_claim(id=1063190)